YOLOv8 Performance Improvement Masterclass

Explore how image size as a hyperparameter affects YOLO v8 accuracy, speed, and mean average precision, with Colab experiments comparing 640 and 320 image sizes on COCO.

Explore how shrinking the image size hyperparameter from 640 to 320 speeds up YOLOv8 video detection by about 1.5x, while reducing accuracy by about 0.088.

Explore how half precision affects YOLO v8 performance, comparing mean average precision on COCO 128 and inference speed. The lecture demonstrates running experiments in Google Colab with a pretrained model.

Explore how half precision speeds up YOLOv8 inference while preserving accuracy, comparing non-half and half-precision runs and showing a speed gain of about 1.3x with minimal accuracy loss.

Accelerate YOLOv8 inference with TensorRT on Nvidia GPUs using Google Colab, converting to TensorRT engine, installing TensorRT, and comparing speed and mean average precision against the original model.

Compare the speed and accuracy of YOLOv8 and TensorRT during video object detection, demonstrate inference with the same notebook, and show a twofold speed gain with minimal accuracy loss.

Explore how to optimize the YOLOv8 model on the CPU using OpenVINO quantization with mixed precision (fp16/fp32), including preparing data, converting to OpenVINO, quantizing in Google Colab, and comparing performance.

Compare YOLOv8 and OpenVINO quantized models for video object detection, measuring mean average precision and speed to reveal performance trade-offs.

Explore the YOLOv8 architecture, including backbone, neck, and head, and learn how convolutional, c2f, bottleneck, spf, and detect blocks enable anchor-free detection across various sizes.

Modify the YOLOv8 architecture for small objects by editing the backbone and head blocks and creating a small-objects architecture file.

Train both the original and modified YOLOv8 models from scratch on a small-objects sperm dataset, showing the 12-block modification speeds training by almost two times with minimal accuracy loss.

Compare the original and modified YOLOv8 models for small objects on video inference, revealing a 1.2x speed increase and minimal accuracy loss from architecture tweaks.

Modify the YOLOv8 architecture to detect medium objects by removing unused blocks and updating connections, creating a dedicated medium object architecture file, and training the model on the sperm dataset.

Learn to modify the YOLOv8 architecture for big objects by deleting unused blocks and configuring depth, width, and max channels with the large variant to boost mAP.

Train from scratch using architecture file; compare original YOLOv8 and a big-objects modified version with 14 blocks on the tomato leaf dataset. The modified version yields higher accuracy (0.756 mAP) vs 0.539 mAP for the original, and takes about 19 minutes for 100 epochs, while the original takes about 15 minutes.

Compare the original and modified YOLOv8 models for big-object detection, showing faster inference and higher accuracy with the modified architecture despite longer training times.

Explore hyperparameters in YOLOv8, including epochs, early stopping patience, batch size, and number of workers, to optimize convergence and mean average precision on your dataset.

Explore how learning rate, momentum, and weight decay affect YOLOv8 training, with experiments showing how default values, warmup, and final learning rate influence mean average precision.

Learn to train the YOLOv8 model using default hyperparameters in a Google Colab workflow, including dataset preparation, data.yaml setup, and running the train command to optimize map.

Train the YOLOv8 model with tuned hyperparameters using the same notebook, improving mean average precision from 0.889 to 0.901 on validation.

Learn how data augmentation enhances YOLOv8 performance, using HSV, degrees, translate, scale, mosaic, and mixup to improve generalization and mean average precision.

Explore data augmentation for YOLOv8, including vertical and horizontal flips, image shift, and perspective, with dataset-guided choices and their impact on mean average precision.

Train the YOLOv8 model using default data augmentation values in a Google Colab notebook, including dataset preparation, setting data.yaml, and a 100-epoch training with resume capability.

Train the YOLOv8 model using recommended data augmentation values to significantly boost mean average precision (mAP) on validation and test sets.

Combine a YOLOv8 architecture modification for big objects with TensorRT to boost accuracy and speed, convert the model to TensorRT, and compare mAP and inference times.

Apply pseudo labeling to speed up YOLOv8 annotations by manually labeling a subset (132 road-sign images), training, then refining model-generated annotations with CVAT-exported YOLO results.

Train a YOLOv8 model on a 132-image road sign dataset using a Google Colab notebook, preparing data, uploading, and splitting into train, val, and test sets.

Train a YOLOv8 model on 132 road sign images, then automatically detect objects in 745 unannotated images, convert detections to YOLO format, and prepare annotations for download.

Train the YOLO model with all images in the road sign dataset to boost detection performance. Mean average precision rises from 0.905 to 0.958, a 0.053 gain.